the ring arrived with cloud transcription plus a local model fallback. the question was how to hook it into the stack, and the app offers an mcp page and a webhooks page. the webhook only receives raw recordings, which is mostly useless for actions. the mcp path works: a home assistant mcp server got created through the config flow api, loaded, and verified by handshake, with light control among its tools.
then the architecture split. a single press is a note: it comes to the agent, gets acted on or answered, and the reply lands on telegram. a real test note about buying tape produced three shopping options with links, prices, and ratings, never a purchase. a double press is an action, and it originally went through the vendor's cloud agent at about five seconds, way too slow for lights, so the prototype pointed the command straight at the local intent engine instead, no llm, no cloud round trip.
the measured result was about four seconds total, and the lights acted before the transcription even appeared as a notification. gesture naming in the app turned out to be slippery: a press that does not register the second click gets labeled as a single press, which confused routing once, so the receiver got upgrades that handle both cases. the rule that finally stuck is simple. every press hits a local command engine first and executes silently in milliseconds, and only notes and questions get the full agent treatment.